| PaperNO | Paper / Abstract |
|
SE6-001
10:50
|
11:10
|
STRUCTURAL HEALTH MONITORING USING MACHINE LEARNING
This paper presents two on-going efforts within the Pacific Earthquake Engineering Research (PEER) center in the area of structural health monitoring (SHM), which applies machine learning (ML) and deep learning (DL) techniques. The first is a data-driven damage assessment, which utilizes ML tools and structural engineering expertise through the so-called Human-Machine Collaboration (H-MC) framework. This framework is applied for damage assessment of instrumented buildings located in the US and Taiwan. One of the damage features identified to be highly efficient is the cumulative absolute velocity (CAV). The second is vision-based automated damage identification and assessment from images. Firstly, we introduce PEER’s effort in open-sourcing the first large-scale and multi-attributes benchmark structural image dataset, namely PEER Hub Image-Net (Φ-Net). Then, based on the Φ-Net framework and dataset, DL techniques are used to train a Structural ImageNet Model (SIM) based on several identification tasks from images, examples of which are the scene level, damage state, structural component type, etc. With well-trained SIMs, they are applied for damage assessment from images collected from 1999 Chi-Chi earthquake and achieved promising results. In summary, the ultimate objective of this study is to realize automated damage detection and severity evaluation for rapid decision making after disasters.
Sifat Muin, Khalid Mosalam, Yuqing Gao
Chi-Chi earthquake, convolutional neural network, cumulative absolute velocity, deep learning, human-machine collaboration, machine learning, structural health monitoring, Φ-net
|
|
SE6-018
11:25
|
11:40
|
A Real Time Seismic Capability Evaluation of School Buildings Using Machine Learning
School buildings have higher structural requirements because they are densely populated infrastructures. Located at circum-pacific seismic belt, the school buildings in Taiwan are under the threats of frequent earthquakes. Once a school building is damaged by earthquake, seismic capability evaluations need to be performed for the possibility of retrofits. The National Center of Research on Earthquake Engineering (NCREE) in Taiwan has developed a procedure for evaluating the seismic capabilities of school buildings. The evaluation is divided into preliminary evaluation and detailed evaluation. Detailed evaluation can provide reliable analysis results. However, it is labor-intensive, time-consuming, and expensive. To compare with detailed evaluation, preliminary evaluation is relatively rapid but also less reliable. To address this problem, we proposed a machine learning-based approach to enhance the performance of prediction. The results showed that the proposed approach can provide reliable predictions on seismic capability evaluations on school buildings.
Nai-Wen Chi, Jyun-Ping Wang, Jia-Hsing Liao, We-Choung Cheng, Chuin-Shan Chen
machine learning, school building, seismic capability evaluation
|
|
SE6-011
11:10
|
11:25
|
USE OF MACHINE LEARNING TECHNIQUES TO DETECT THE LOCATIONS OF EARTHQUAKE-INDUCED SLOPE FAILURES
The earthquake-induced slope failures caused a large number of human casualties during recent earthquakes in Japan. The 2016 Kumamoto and 2018 Hokkaido Eastern Iburi earthquakes triggered hundreds of slope failures, and resulted in damage to buildings and infrastructures. Based on the background, the objective of this study is to develop a numerical model to detect the locations of slope failures because of an earthquake. To achieve the objective, machine learning techniques are employed in this study. The accuracy is evaluated quantitatively based on diagnostic test indicators, and a suitable numerical model is proposed by this study.
Yoshihisa Maruyama
2016 Kumamoto earthquake, diagnostic test indicator, earthquake-induced slope failure, machine learning
|
|
SE6-015
11:40
|
11:55
|
ONLINE MODEL UPDATING FOR THE ADVANCED HYBRID SIMULATIONS OF A STEEL PANEL DAMPER SUBSTRUCTURE
In a hybrid simulation, the target structure under investigation can be partitioned into multiple substructures. These substructures can be divided into two categories: numerical substructure (NS) and physical substructure (PS). The hybrid simulation allows the numerical and physical substructures to be integrated such that the interaction between the substructures can be taken into account in seismic performance assessment. As a result, the hybrid simulation can offer a cost-effective alternative to the shaking table test while a full-scale target structure is investigated. However, the application of conventional hybrid simulation is frequently restricted due to the limited numbers of facilities and specimens. It is inevitable to numerically represent several structural elements, that are similar to or the same as the PSs, in the conventional hybrid simulations. Thus, the advantage of hybrid simulation would be diminished due to the inaccurate modeling of these NSs. In order to address the aforementioned issue, the researchers at the Taiwan National Center for Research on Earthquake Engineering (NCREE) have developed the techniques of online model updating (OMU) to take up the mentioned challenge in hybrid simulation. The NCREE researchers proposed the gradient-based parameter identification (PI) method for OMU. In this study, the proposed PI method is applied to the OMU schemes of the hybrid simulations of a steel panel damper (SPD) substructure conducted using a multi-axial testing system at the NCREE in 2017. The structure under investigation is a three-dimensional six-story moment resisting frame (MRF) with four SPDs installed at each story. In this advanced hybrid simulations with OMU, only one SPD is represented as the PS, namely the SPD specimen. The rest of the SPDs and MRF are represented as the NSs. It was found that through the OMU, the proper parameter values of the constitutive model representing the experimentally measured force - deformation relationships of the PS can be identified effectively. Using the identified parameter values, the constitutive models of relevant SPD elements in the NS can be online rectified during the hybrid simulations. Hence, the accuracy of the hybrid simulations can be improved. The test results verified and demonstrated the effectiveness and benefits of the OMU with the proposed PI method for the advanced hybrid simulations.
Ming-Chieh Chuang, Kung-Juin Wang, Keh-Chyuan Tsai
hybrid simulation, model updating, parameter identification, steel panel damper
|
|
SE6-016
11:55
|
12:10
|
Methodology for Earthquake-Fire Coupled Hybrid Simulation
Fire often occurs during or sequentially after major earthquakes. Such extreme earthquake-fire coupled events inflicted heavy casualties and economic losses during the 1906 San Francisco earthquake and 1994 Northridge earthquake in the US, and 1995 Kobe earthquake in Japan. With increasingly dense urban development of major cities around the world, improvements to the safety and resilience of structures against earthquake-fire multi-hazard events is crucial.To design structures for the multi-hazard scenario of fire following earthquake, it is necessary to have information not only about the material properties and behaviour of structures subjected to both mechanical and thermal load actions, but also about the sequence of the load history and the progression of damage in the sequential multi-hazard event. To evaluate the system- level performance of structures exposed to the sequential loads of earthquake and fire in a practical and cost-effective manner, the hybrid simulation technique that combines physical testing and numerical modelling, previously recently developed for earthquake hazards, is further developed in this research for application to fire following earthquake multi-hazard events.This paper presents a framework for real-time multi-hazard (post-earthquake fire) hybrid simulation of full-scale specimens that considers full interaction between the thermal and mechanical behaviour of the structure. In the proposed framework, the element of the prototype structure that is exposed to the sequence of loads is selected as the physical specimen (physical domain) while the remainder structure is numerically modelled (numerical domain). New OpenFresco objects for beams/columns are developed to include both temperature and mechanical degrees-of-freedom with full compatibility of deformations as well as thermal flux and force equilibrium at the interface between the physical and numerical domains in real-time. A illustrative example of the new fire following earthquake hybrid simulation framework is also presented to demonstrate the validity of the methodology.
David Lau, Jeffrey Erochko, Zhimeng Yu, Oh-Sung Kwon, Ahmed Kashef
fire following earthquake, hybrid simulation, multi-hazards, performance-based design
|